ShinojiResearch/Senku-70B-Full

TEXT GENERATIONConcurrency Cost:4Model Size:69BQuant:FP8Ctx Length:32kPublished:Feb 6, 2024License:cc0-1.0Architecture:Transformer0.1K Open Weights Cold

ShinojiResearch/Senku-70B-Full is a 69 billion parameter language model fine-tuned from a dequantized version of miqu-70b-sf, which is based on an alleged early Mistral-70B. This model is specifically fine-tuned on the Slimorca dataset and achieves an EQ-Bench score of 85.09 using the ChatML prompt format. It is notable for being the first open-weight model to surpass GPT-4 on EQ-Bench, making it suitable for complex reasoning and general conversational AI tasks.

Loading preview...

ShinojiResearch/Senku-70B-Full Overview

Senku-70B-Full is a 69 billion parameter language model developed by ShinojiResearch. It is a fine-tuned version of miqu-70b-sf, which itself is a dequantized variant of an alleged early Mistral-70B model. The model was trained using the Axolotl framework on the Slimorca dataset, with a sequence length of 8192 tokens and a learning rate of 0.0002.

Key Capabilities & Performance

  • Exceptional Reasoning: Achieves an impressive 85.09 EQ-Bench score using the ChatML prompt template, notably dethroning GPT-4 on this benchmark. This indicates strong performance in complex reasoning tasks.
  • Robust General Performance: On the Open LLM Leaderboard, it demonstrates solid average performance with a score of 75.44. Specific benchmark results include:
    • AI2 Reasoning Challenge (25-Shot): 71.50
    • HellaSwag (10-Shot): 87.88
    • MMLU (5-Shot): 75.20
    • GSM8k (5-Shot): 71.34
  • Optimized Prompting: The model performs best with the ChatML prompt format, which also resolves a known stop token issue present in the base Miqu dequant model.

When to Use This Model

  • Advanced Reasoning Applications: Ideal for use cases requiring high-level reasoning and problem-solving, as evidenced by its leading EQ-Bench score.
  • General Conversational AI: Its strong performance across various benchmarks makes it suitable for a wide range of conversational and instruction-following tasks.
  • Research and Development: Offers a powerful base for further fine-tuning or experimentation, especially given its lineage from an alleged Mistral-70B variant.